Computer Simulation of Articulatory Activity in Speech Production

Paul MermelsteinIJCAI 1969

Introduction

Traditional speech synthesis in the late 1960s primarily focused on the acoustic domain, attempting to replicate resonant frequencies (formants) directly. However, this approach often failed to capture the complex, non-linear relationship between physical anatomy and sound. P. Mermelstein’s work at Bell Telephone Laboratories introduced a shift toward articulatory modeling, simulating the physical movement of the human vocal tract to produce more natural and intelligible speech.

Modeling the Vocal Tract

The core of the simulation is a representation of the vocal tract as a non-uniform, variable acoustic transmission line. The tract is divided into discrete sections, typically 1.5 cm in length, defined by their cross-sectional areas. This physical grounding allows the system to account for the constraints of human musculature and anatomy.

Block diagram of the articulatory speech production model, showing the nasal, pharyngeal, and oral sections along with excitation sources.

As shown in the block diagram above, the model incorporates several key components:

  • Excitation Sources: A periodic source simulates vocal cord vibration for voiced sounds, while a noise generator provides the turbulence necessary for fricatives.
  • Nasal Coupling: A secondary transmission line represents the nasal passage, allowing for the synthesis of nasalized vowels and consonants.
  • Propagation Units: Detailed in the inset, these units handle the reflection and transmission of acoustic waves between sections based on area changes.

Dynamic Synthesis and Coarticulation

A major insight of this research is that natural speech requires constant motion. Static articulatory shapes produce artificial, "robotic" sounds. To address this, the system uses linear interpolation of reflection coefficients between target articulatory states, creating fluid transitions that mimic biological movement.

This articulatory approach simplifies the problem of coarticulation—where the sound of a phoneme is influenced by its neighbors. By modeling the positions of the tongue, lips, and jaw, the system can apply predictable physical rules to generate the complex frequency shifts observed in natural speech.

Results and Evaluation

The system's effectiveness was demonstrated through the synthesis of the sentence, "This is a story about a man." Using a sequence of 27 articulatory targets to represent 19 phonemes, the simulation produced highly intelligible speech.

Spectrogram showing the resonant frequency trajectories of the synthesized speech.

Qualitative analysis of the resulting spectrograms (see above) revealed that the machine-generated resonant frequency trajectories closely matched those of natural human speech. While the simulation ran at approximately 40 times real-time—a significant computational burden for the era—it proved the viability of articulatory-based synthesis.

Conclusion and Future Directions

Mermelstein’s simulation highlighted that acoustic complexity often arises from relatively simple physical movements. While the model had limitations, such as manual tuning requirements and simplified physics (e.g., hard-walled tube assumptions), it laid the groundwork for future automated articulatory rules and more sophisticated physiological models. Future work aimed to reduce computational lag and refine the modeling of complex sounds like /l/ and /r/.

Summary updated

BibTeX
@inproceedings{mermelstein1969computer,
  author = {Paul Mermelstein},
  title = {{Computer Simulation of Articulatory Activity in Speech Production}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1969},
}
Figure